Merge "external/webp: update to v1.1.0-28-g55a080e5" am: c4147d8f00 am: 663b190f19

Original change: https://android-review.googlesource.com/c/platform/external/webp/+/1425968

Change-Id: I1bddaa080ae7435545511850c7e181007e8aac6e

24 files changed, 496 insertions, 400 deletions

diff --git a/README b/README
index 05927276..0baba6ab 100644
--- a/README
+++ b/README
@@ -254,6 +254,8 @@ Options:
   -partition_limit <int> . limit quality to fit the 512k limit on
                            the first partition (0=no degradation ... 100=full)
   -pass <int> ............ analysis pass number (1..10)
+  -qrange <min> <max> .... specifies the permissible quality range
+                           (default: 0 100)
   -crop <x> <y> <w> <h> .. crop picture with the given rectangle
   -resize <w> <h> ........ resize picture (after any cropping)
   -mt .................... use multi-threading if available
@@ -294,6 +296,7 @@ Experimental Options:
   -af .................... auto-adjust filter strength
   -pre <int> ............. pre-processing filter
 
+
 The main options you might want to try in order to further tune the
 visual quality are:
  -preset
diff --git a/README.android b/README.android
index 94fd4d4a..4c9394c3 100644
--- a/README.android
+++ b/README.android
@@ -1,5 +1,5 @@
 URL: https://chromium.googlesource.com/webm/libwebp
-Version: v1.1.0
+Version: v1.1.0-28-g55a080e5
 License: Google BSD like
 
 Local modifications:
diff --git a/README.version b/README.version
index f94ac059..32ba7bd5 100644
--- a/README.version
+++ b/README.version
@@ -1,3 +1,3 @@
-URL: https://storage.googleapis.com/downloads.webmproject.org/releases/webp/libwebp-1.1.0.tar.gz
-Version: 1.1.0
+URL: https://chromium.googlesource.com/webm/libwebp/+archive/v1.1.0-28-g55a080e5.tar.gz
+Version: v1.1.0-28-g55a080e5
 BugComponent: 20174
diff --git a/include/webp/encode.h b/include/webp/encode.h
index 655166e7..54938dfe 100644
--- a/include/webp/encode.h
+++ b/include/webp/encode.h
@@ -20,7 +20,7 @@
 extern "C" {
 #endif
 
-#define WEBP_ENCODER_ABI_VERSION 0x020f    // MAJOR(8b) + MINOR(8b)
+#define WEBP_ENCODER_ABI_VERSION 0x0210    // MAJOR(8b) + MINOR(8b)
 
 // Note: forward declaring enumerations is not allowed in (strict) C and C++,
 // the types are left here for reference.
@@ -148,7 +148,8 @@ struct WebPConfig {
   int use_delta_palette;  // reserved for future lossless feature
   int use_sharp_yuv;      // if needed, use sharp (and slow) RGB->YUV conversion
 
-  uint32_t pad[2];        // padding for later use
+  int qmin;               // minimum permissible quality factor
+  int qmax;               // maximum permissible quality factor
 };
 
 // Enumerate some predefined settings for WebPConfig, depending on the type
diff --git a/src/dec/io_dec.c b/src/dec/io_dec.c
index e603f19c..727cce7b 100644
--- a/src/dec/io_dec.c
+++ b/src/dec/io_dec.c
@@ -25,21 +25,16 @@
 static int EmitYUV(const VP8Io* const io, WebPDecParams* const p) {
   WebPDecBuffer* output = p->output;
   const WebPYUVABuffer* const buf = &output->u.YUVA;
-  uint8_t* const y_dst = buf->y + io->mb_y * buf->y_stride;
-  uint8_t* const u_dst = buf->u + (io->mb_y >> 1) * buf->u_stride;
-  uint8_t* const v_dst = buf->v + (io->mb_y >> 1) * buf->v_stride;
+  uint8_t* const y_dst = buf->y + (size_t)io->mb_y * buf->y_stride;
+  uint8_t* const u_dst = buf->u + (size_t)(io->mb_y >> 1) * buf->u_stride;
+  uint8_t* const v_dst = buf->v + (size_t)(io->mb_y >> 1) * buf->v_stride;
   const int mb_w = io->mb_w;
   const int mb_h = io->mb_h;
   const int uv_w = (mb_w + 1) / 2;
   const int uv_h = (mb_h + 1) / 2;
-  int j;
-  for (j = 0; j < mb_h; ++j) {
-    memcpy(y_dst + j * buf->y_stride, io->y + j * io->y_stride, mb_w);
-  }
-  for (j = 0; j < uv_h; ++j) {
-    memcpy(u_dst + j * buf->u_stride, io->u + j * io->uv_stride, uv_w);
-    memcpy(v_dst + j * buf->v_stride, io->v + j * io->uv_stride, uv_w);
-  }
+  WebPCopyPlane(io->y, io->y_stride, y_dst, buf->y_stride, mb_w, mb_h);
+  WebPCopyPlane(io->u, io->uv_stride, u_dst, buf->u_stride, uv_w, uv_h);
+  WebPCopyPlane(io->v, io->uv_stride, v_dst, buf->v_stride, uv_w, uv_h);
   return io->mb_h;
 }
 
@@ -47,7 +42,7 @@ static int EmitYUV(const VP8Io* const io, WebPDecParams* const p) {
 static int EmitSampledRGB(const VP8Io* const io, WebPDecParams* const p) {
   WebPDecBuffer* const output = p->output;
   WebPRGBABuffer* const buf = &output->u.RGBA;
-  uint8_t* const dst = buf->rgba + io->mb_y * buf->stride;
+  uint8_t* const dst = buf->rgba + (size_t)io->mb_y * buf->stride;
   WebPSamplerProcessPlane(io->y, io->y_stride,
                           io->u, io->v, io->uv_stride,
                           dst, buf->stride, io->mb_w, io->mb_h,
@@ -62,7 +57,7 @@ static int EmitSampledRGB(const VP8Io* const io, WebPDecParams* const p) {
 static int EmitFancyRGB(const VP8Io* const io, WebPDecParams* const p) {
   int num_lines_out = io->mb_h;   // a priori guess
   const WebPRGBABuffer* const buf = &p->output->u.RGBA;
-  uint8_t* dst = buf->rgba + io->mb_y * buf->stride;
+  uint8_t* dst = buf->rgba + (size_t)io->mb_y * buf->stride;
   WebPUpsampleLinePairFunc upsample = WebPUpsamplers[p->output->colorspace];
   const uint8_t* cur_y = io->y;
   const uint8_t* cur_u = io->u;
@@ -133,7 +128,7 @@ static int EmitAlphaYUV(const VP8Io* const io, WebPDecParams* const p,
   const WebPYUVABuffer* const buf = &p->output->u.YUVA;
   const int mb_w = io->mb_w;
   const int mb_h = io->mb_h;
-  uint8_t* dst = buf->a + io->mb_y * buf->a_stride;
+  uint8_t* dst = buf->a + (size_t)io->mb_y * buf->a_stride;
   int j;
   (void)expected_num_lines_out;
   assert(expected_num_lines_out == mb_h);
@@ -186,7 +181,7 @@ static int EmitAlphaRGB(const VP8Io* const io, WebPDecParams* const p,
         (colorspace == MODE_ARGB || colorspace == MODE_Argb);
     const WebPRGBABuffer* const buf = &p->output->u.RGBA;
     int num_rows;
-    const int start_y = GetAlphaSourceRow(io, &alpha, &num_rows);
+    const size_t start_y = GetAlphaSourceRow(io, &alpha, &num_rows);
     uint8_t* const base_rgba = buf->rgba + start_y * buf->stride;
     uint8_t* const dst = base_rgba + (alpha_first ? 0 : 3);
     const int has_alpha = WebPDispatchAlpha(alpha, io->width, mb_w,
@@ -210,7 +205,7 @@ static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p,
     const WEBP_CSP_MODE colorspace = p->output->colorspace;
     const WebPRGBABuffer* const buf = &p->output->u.RGBA;
     int num_rows;
-    const int start_y = GetAlphaSourceRow(io, &alpha, &num_rows);
+    const size_t start_y = GetAlphaSourceRow(io, &alpha, &num_rows);
     uint8_t* const base_rgba = buf->rgba + start_y * buf->stride;
 #if (WEBP_SWAP_16BIT_CSP == 1)
     uint8_t* alpha_dst = base_rgba;
@@ -276,9 +271,9 @@ static int EmitRescaledYUV(const VP8Io* const io, WebPDecParams* const p) {
 static int EmitRescaledAlphaYUV(const VP8Io* const io, WebPDecParams* const p,
                                 int expected_num_lines_out) {
   const WebPYUVABuffer* const buf = &p->output->u.YUVA;
-  uint8_t* const dst_a = buf->a + p->last_y * buf->a_stride;
+  uint8_t* const dst_a = buf->a + (size_t)p->last_y * buf->a_stride;
   if (io->a != NULL) {
-    uint8_t* const dst_y = buf->y + p->last_y * buf->y_stride;
+    uint8_t* const dst_y = buf->y + (size_t)p->last_y * buf->y_stride;
     const int num_lines_out = Rescale(io->a, io->width, io->mb_h, p->scaler_a);
     assert(expected_num_lines_out == num_lines_out);
     if (num_lines_out > 0) {   // unmultiply the Y
@@ -356,7 +351,7 @@ static int ExportRGB(WebPDecParams* const p, int y_pos) {
   const WebPYUV444Converter convert =
       WebPYUV444Converters[p->output->colorspace];
   const WebPRGBABuffer* const buf = &p->output->u.RGBA;
-  uint8_t* dst = buf->rgba + y_pos * buf->stride;
+  uint8_t* dst = buf->rgba + (size_t)y_pos * buf->stride;
   int num_lines_out = 0;
   // For RGB rescaling, because of the YUV420, current scan position
   // U/V can be +1/-1 line from the Y one.  Hence the double test.
@@ -383,15 +378,15 @@ static int EmitRescaledRGB(const VP8Io* const io, WebPDecParams* const p) {
   while (j < mb_h) {
     const int y_lines_in =
         WebPRescalerImport(p->scaler_y, mb_h - j,
-                           io->y + j * io->y_stride, io->y_stride);
+                           io->y + (size_t)j * io->y_stride, io->y_stride);
     j += y_lines_in;
     if (WebPRescaleNeededLines(p->scaler_u, uv_mb_h - uv_j)) {
-      const int u_lines_in =
-          WebPRescalerImport(p->scaler_u, uv_mb_h - uv_j,
-                             io->u + uv_j * io->uv_stride, io->uv_stride);
-      const int v_lines_in =
-          WebPRescalerImport(p->scaler_v, uv_mb_h - uv_j,
-                             io->v + uv_j * io->uv_stride, io->uv_stride);
+      const int u_lines_in = WebPRescalerImport(
+          p->scaler_u, uv_mb_h - uv_j, io->u + (size_t)uv_j * io->uv_stride,
+          io->uv_stride);
+      const int v_lines_in = WebPRescalerImport(
+          p->scaler_v, uv_mb_h - uv_j, io->v + (size_t)uv_j * io->uv_stride,
+          io->uv_stride);
       (void)v_lines_in;   // remove a gcc warning
       assert(u_lines_in == v_lines_in);
       uv_j += u_lines_in;
@@ -403,7 +398,7 @@ static int EmitRescaledRGB(const VP8Io* const io, WebPDecParams* const p) {
 
 static int ExportAlpha(WebPDecParams* const p, int y_pos, int max_lines_out) {
   const WebPRGBABuffer* const buf = &p->output->u.RGBA;
-  uint8_t* const base_rgba = buf->rgba + y_pos * buf->stride;
+  uint8_t* const base_rgba = buf->rgba + (size_t)y_pos * buf->stride;
   const WEBP_CSP_MODE colorspace = p->output->colorspace;
   const int alpha_first =
       (colorspace == MODE_ARGB || colorspace == MODE_Argb);
@@ -431,7 +426,7 @@ static int ExportAlpha(WebPDecParams* const p, int y_pos, int max_lines_out) {
 static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos,
                                int max_lines_out) {
   const WebPRGBABuffer* const buf = &p->output->u.RGBA;
-  uint8_t* const base_rgba = buf->rgba + y_pos * buf->stride;
+  uint8_t* const base_rgba = buf->rgba + (size_t)y_pos * buf->stride;
 #if (WEBP_SWAP_16BIT_CSP == 1)
   uint8_t* alpha_dst = base_rgba;
 #else
@@ -470,7 +465,7 @@ static int EmitRescaledAlphaRGB(const VP8Io* const io, WebPDecParams* const p,
     int lines_left = expected_num_out_lines;
     const int y_end = p->last_y + lines_left;
     while (lines_left > 0) {
-      const int row_offset = scaler->src_y - io->mb_y;
+      const size_t row_offset = scaler->src_y - io->mb_y;
       WebPRescalerImport(scaler, io->mb_h + io->mb_y - scaler->src_y,
                          io->a + row_offset * io->width, io->width);
       lines_left -= p->emit_alpha_row(p, y_end - lines_left, lines_left);
diff --git a/src/demux/anim_decode.c b/src/demux/anim_decode.c
index 05dd7073..3dcacc35 100644
--- a/src/demux/anim_decode.c
+++ b/src/demux/anim_decode.c
@@ -346,12 +346,15 @@ int WebPAnimDecoderGetNext(WebPAnimDecoder* dec,
   {
     const uint8_t* in = iter.fragment.bytes;
     const size_t in_size = iter.fragment.size;
-    const size_t out_offset =
-        (iter.y_offset * width + iter.x_offset) * NUM_CHANNELS;
+    const uint32_t stride = width * NUM_CHANNELS;  // at most 25 + 2 bits
+    const uint64_t out_offset = (uint64_t)iter.y_offset * stride +
+                                (uint64_t)iter.x_offset * NUM_CHANNELS;  // 53b
+    const uint64_t size = (uint64_t)iter.height * stride;  // at most 25 + 27b
     WebPDecoderConfig* const config = &dec->config_;
     WebPRGBABuffer* const buf = &config->output.u.RGBA;
-    buf->stride = NUM_CHANNELS * width;
-    buf->size = buf->stride * iter.height;
+    if ((size_t)size != size) goto Error;
+    buf->stride = (int)stride;
+    buf->size = (size_t)size;
     buf->rgba = dec->curr_frame_ + out_offset;
 
     if (WebPDecode(in, in_size, config) != VP8_STATUS_OK) {
diff --git a/src/dsp/alpha_processing.c b/src/dsp/alpha_processing.c
index 819d1391..3a27990d 100644
--- a/src/dsp/alpha_processing.c
+++ b/src/dsp/alpha_processing.c
@@ -359,6 +359,11 @@ static int HasAlpha32b_C(const uint8_t* src, int length) {
   return 0;
 }
 
+static void AlphaReplace_C(uint32_t* src, int length, uint32_t color) {
+  int x;
+  for (x = 0; x < length; ++x) if ((src[x] >> 24) == 0) src[x] = color;
+}
+
 //------------------------------------------------------------------------------
 // Simple channel manipulations.
 
@@ -400,6 +405,7 @@ void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b,
 
 int (*WebPHasAlpha8b)(const uint8_t* src, int length);
 int (*WebPHasAlpha32b)(const uint8_t* src, int length);
+void (*WebPAlphaReplace)(uint32_t* src, int length, uint32_t color);
 
 //------------------------------------------------------------------------------
 // Init function
@@ -428,6 +434,7 @@ WEBP_DSP_INIT_FUNC(WebPInitAlphaProcessing) {
 
   WebPHasAlpha8b = HasAlpha8b_C;
   WebPHasAlpha32b = HasAlpha32b_C;
+  WebPAlphaReplace = AlphaReplace_C;
 
   // If defined, use CPUInfo() to overwrite some pointers with faster versions.
   if (VP8GetCPUInfo != NULL) {
@@ -469,4 +476,5 @@ WEBP_DSP_INIT_FUNC(WebPInitAlphaProcessing) {
   assert(WebPPackRGB != NULL);
   assert(WebPHasAlpha8b != NULL);
   assert(WebPHasAlpha32b != NULL);
+  assert(WebPAlphaReplace != NULL);
 }
diff --git a/src/dsp/alpha_processing_sse2.c b/src/dsp/alpha_processing_sse2.c
index 2871c56d..f6c6e0fb 100644
--- a/src/dsp/alpha_processing_sse2.c
+++ b/src/dsp/alpha_processing_sse2.c
@@ -265,6 +265,27 @@ static int HasAlpha32b_SSE2(const uint8_t* src, int length) {
   return 0;
 }
 
+static void AlphaReplace_SSE2(uint32_t* src, int length, uint32_t color) {
+  const __m128i m_color = _mm_set1_epi32(color);
+  const __m128i zero = _mm_setzero_si128();
+  int i = 0;
+  for (; i + 8 <= length; i += 8) {
+    const __m128i a0 = _mm_loadu_si128((const __m128i*)(src + i + 0));
+    const __m128i a1 = _mm_loadu_si128((const __m128i*)(src + i + 4));
+    const __m128i b0 = _mm_srai_epi32(a0, 24);
+    const __m128i b1 = _mm_srai_epi32(a1, 24);
+    const __m128i c0 = _mm_cmpeq_epi32(b0, zero);
+    const __m128i c1 = _mm_cmpeq_epi32(b1, zero);
+    const __m128i d0 = _mm_and_si128(c0, m_color);
+    const __m128i d1 = _mm_and_si128(c1, m_color);
+    const __m128i e0 = _mm_andnot_si128(c0, a0);
+    const __m128i e1 = _mm_andnot_si128(c1, a1);
+    _mm_storeu_si128((__m128i*)(src + i + 0), _mm_or_si128(d0, e0));
+    _mm_storeu_si128((__m128i*)(src + i + 4), _mm_or_si128(d1, e1));
+  }
+  for (; i < length; ++i) if ((src[i] >> 24) == 0) src[i] = color;
+}
+
 // -----------------------------------------------------------------------------
 // Apply alpha value to rows
 
@@ -334,6 +355,7 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingSSE2(void) {
 
   WebPHasAlpha8b = HasAlpha8b_SSE2;
   WebPHasAlpha32b = HasAlpha32b_SSE2;
+  WebPAlphaReplace = AlphaReplace_SSE2;
 }
 
 #else  // !WEBP_USE_SSE2
diff --git a/src/dsp/dec_neon.c b/src/dsp/dec_neon.c
index 239ec416..fa851707 100644
--- a/src/dsp/dec_neon.c
+++ b/src/dsp/dec_neon.c
@@ -1283,12 +1283,12 @@ static void DC4_NEON(uint8_t* dst) {    // DC
   const uint8x8_t A = vld1_u8(dst - BPS);  // top row
   const uint16x4_t p0 = vpaddl_u8(A);  // cascading summation of the top
   const uint16x4_t p1 = vpadd_u16(p0, p0);
-  const uint16x8_t L0 = vmovl_u8(vld1_u8(dst + 0 * BPS - 1));
-  const uint16x8_t L1 = vmovl_u8(vld1_u8(dst + 1 * BPS - 1));
-  const uint16x8_t L2 = vmovl_u8(vld1_u8(dst + 2 * BPS - 1));
-  const uint16x8_t L3 = vmovl_u8(vld1_u8(dst + 3 * BPS - 1));
-  const uint16x8_t s0 = vaddq_u16(L0, L1);
-  const uint16x8_t s1 = vaddq_u16(L2, L3);
+  const uint8x8_t L0 = vld1_u8(dst + 0 * BPS - 1);
+  const uint8x8_t L1 = vld1_u8(dst + 1 * BPS - 1);
+  const uint8x8_t L2 = vld1_u8(dst + 2 * BPS - 1);
+  const uint8x8_t L3 = vld1_u8(dst + 3 * BPS - 1);
+  const uint16x8_t s0 = vaddl_u8(L0, L1);
+  const uint16x8_t s1 = vaddl_u8(L2, L3);
   const uint16x8_t s01 = vaddq_u16(s0, s1);
   const uint16x8_t sum = vaddq_u16(s01, vcombine_u16(p1, p1));
   const uint8x8_t dc0 = vrshrn_n_u16(sum, 3);  // (sum + 4) >> 3
@@ -1429,8 +1429,7 @@ static WEBP_INLINE void DC8_NEON(uint8_t* dst, int do_top, int do_left) {
   if (do_top) {
     const uint8x8_t A = vld1_u8(dst - BPS);  // top row
 #if defined(__aarch64__)
-    const uint16x8_t B = vmovl_u8(A);
-    const uint16_t p2 = vaddvq_u16(B);
+    const uint16_t p2 = vaddlv_u8(A);
     sum_top = vdupq_n_u16(p2);
 #else
     const uint16x4_t p0 = vpaddl_u8(A);  // cascading summation of the top
@@ -1441,18 +1440,18 @@ static WEBP_INLINE void DC8_NEON(uint8_t* dst, int do_top, int do_left) {
   }
 
   if (do_left) {
-    const uint16x8_t L0 = vmovl_u8(vld1_u8(dst + 0 * BPS - 1));
-    const uint16x8_t L1 = vmovl_u8(vld1_u8(dst + 1 * BPS - 1));
-    const uint16x8_t L2 = vmovl_u8(vld1_u8(dst + 2 * BPS - 1));
-    const uint16x8_t L3 = vmovl_u8(vld1_u8(dst + 3 * BPS - 1));
-    const uint16x8_t L4 = vmovl_u8(vld1_u8(dst + 4 * BPS - 1));
-    const uint16x8_t L5 = vmovl_u8(vld1_u8(dst + 5 * BPS - 1));
-    const uint16x8_t L6 = vmovl_u8(vld1_u8(dst + 6 * BPS - 1));
-    const uint16x8_t L7 = vmovl_u8(vld1_u8(dst + 7 * BPS - 1));
-    const uint16x8_t s0 = vaddq_u16(L0, L1);
-    const uint16x8_t s1 = vaddq_u16(L2, L3);
-    const uint16x8_t s2 = vaddq_u16(L4, L5);
-    const uint16x8_t s3 = vaddq_u16(L6, L7);
+    const uint8x8_t L0 = vld1_u8(dst + 0 * BPS - 1);
+    const uint8x8_t L1 = vld1_u8(dst + 1 * BPS - 1);
+    const uint8x8_t L2 = vld1_u8(dst + 2 * BPS - 1);
+    const uint8x8_t L3 = vld1_u8(dst + 3 * BPS - 1);
+    const uint8x8_t L4 = vld1_u8(dst + 4 * BPS - 1);
+    const uint8x8_t L5 = vld1_u8(dst + 5 * BPS - 1);
+    const uint8x8_t L6 = vld1_u8(dst + 6 * BPS - 1);
+    const uint8x8_t L7 = vld1_u8(dst + 7 * BPS - 1);
+    const uint16x8_t s0 = vaddl_u8(L0, L1);
+    const uint16x8_t s1 = vaddl_u8(L2, L3);
+    const uint16x8_t s2 = vaddl_u8(L4, L5);
+    const uint16x8_t s3 = vaddl_u8(L6, L7);
     const uint16x8_t s01 = vaddq_u16(s0, s1);
     const uint16x8_t s23 = vaddq_u16(s2, s3);
     sum_left = vaddq_u16(s01, s23);
@@ -1512,29 +1511,34 @@ static WEBP_INLINE void DC16_NEON(uint8_t* dst, int do_top, int do_left) {
 
   if (do_top) {
     const uint8x16_t A = vld1q_u8(dst - BPS);  // top row
+#if defined(__aarch64__)
+    const uint16_t p3 = vaddlvq_u8(A);
+    sum_top = vdupq_n_u16(p3);
+#else
     const uint16x8_t p0 = vpaddlq_u8(A);  // cascading summation of the top
     const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
     const uint16x4_t p2 = vpadd_u16(p1, p1);
     const uint16x4_t p3 = vpadd_u16(p2, p2);
     sum_top = vcombine_u16(p3, p3);
+#endif
   }
 
   if (do_left) {
     int i;
     sum_left = vdupq_n_u16(0);
     for (i = 0; i < 16; i += 8) {
-      const uint16x8_t L0 = vmovl_u8(vld1_u8(dst + (i + 0) * BPS - 1));
-      const uint16x8_t L1 = vmovl_u8(vld1_u8(dst + (i + 1) * BPS - 1));
-      const uint16x8_t L2 = vmovl_u8(vld1_u8(dst + (i + 2) * BPS - 1));
-      const uint16x8_t L3 = vmovl_u8(vld1_u8(dst + (i + 3) * BPS - 1));
-      const uint16x8_t L4 = vmovl_u8(vld1_u8(dst + (i + 4) * BPS - 1));
-      const uint16x8_t L5 = vmovl_u8(vld1_u8(dst + (i + 5) * BPS - 1));
-      const uint16x8_t L6 = vmovl_u8(vld1_u8(dst + (i + 6) * BPS - 1));
-      const uint16x8_t L7 = vmovl_u8(vld1_u8(dst + (i + 7) * BPS - 1));
-      const uint16x8_t s0 = vaddq_u16(L0, L1);
-      const uint16x8_t s1 = vaddq_u16(L2, L3);
-      const uint16x8_t s2 = vaddq_u16(L4, L5);
-      const uint16x8_t s3 = vaddq_u16(L6, L7);
+      const uint8x8_t L0 = vld1_u8(dst + (i + 0) * BPS - 1);
+      const uint8x8_t L1 = vld1_u8(dst + (i + 1) * BPS - 1);
+      const uint8x8_t L2 = vld1_u8(dst + (i + 2) * BPS - 1);
+      const uint8x8_t L3 = vld1_u8(dst + (i + 3) * BPS - 1);
+      const uint8x8_t L4 = vld1_u8(dst + (i + 4) * BPS - 1);
+      const uint8x8_t L5 = vld1_u8(dst + (i + 5) * BPS - 1);
+      const uint8x8_t L6 = vld1_u8(dst + (i + 6) * BPS - 1);
+      const uint8x8_t L7 = vld1_u8(dst + (i + 7) * BPS - 1);
+      const uint16x8_t s0 = vaddl_u8(L0, L1);
+      const uint16x8_t s1 = vaddl_u8(L2, L3);
+      const uint16x8_t s2 = vaddl_u8(L4, L5);
+      const uint16x8_t s3 = vaddl_u8(L6, L7);
       const uint16x8_t s01 = vaddq_u16(s0, s1);
       const uint16x8_t s23 = vaddq_u16(s2, s3);
       const uint16x8_t sum = vaddq_u16(s01, s23);
diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index a784de33..cf50bb13 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -193,6 +193,12 @@ extern "C" {
 #endif
 #endif
 
+// If 'ptr' is NULL, returns NULL. Otherwise returns 'ptr + off'.
+// Prevents undefined behavior sanitizer nullptr-with-nonzero-offset warning.
+#if !defined(WEBP_OFFSET_PTR)
+#define WEBP_OFFSET_PTR(ptr, off) (((ptr) == NULL) ? NULL : ((ptr) + (off)))
+#endif
+
 // Regularize the definition of WEBP_SWAP_16BIT_CSP (backward compatibility)
 #if !defined(WEBP_SWAP_16BIT_CSP)
 #define WEBP_SWAP_16BIT_CSP 0
@@ -632,6 +638,8 @@ extern void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b,
 extern int (*WebPHasAlpha8b)(const uint8_t* src, int length);
 // This function returns true if src[4*i] contains a value different from 0xff.
 extern int (*WebPHasAlpha32b)(const uint8_t* src, int length);
+// replaces transparent values in src[] by 'color'.
+extern void (*WebPAlphaReplace)(uint32_t* src, int length, uint32_t color);
 
 // To be called first before using the above.
 void WebPInitAlphaProcessing(void);
diff --git a/src/dsp/lossless_enc_sse2.c b/src/dsp/lossless_enc_sse2.c
index e676f6fd..68d1ac3c 100644
--- a/src/dsp/lossless_enc_sse2.c
+++ b/src/dsp/lossless_enc_sse2.c
@@ -460,20 +460,22 @@ static void PredictorSub0_SSE2(const uint32_t* in, const uint32_t* upper,
   (void)upper;
 }
 
-#define GENERATE_PREDICTOR_1(X, IN)                                           \
-static void PredictorSub##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
-                                   int num_pixels, uint32_t* out) {           \
-  int i;                                                                      \
-  for (i = 0; i + 4 <= num_pixels; i += 4) {                                  \
-    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);              \
-    const __m128i pred = _mm_loadu_si128((const __m128i*)&(IN));              \
-    const __m128i res = _mm_sub_epi8(src, pred);                              \
-    _mm_storeu_si128((__m128i*)&out[i], res);                                 \
-  }                                                                           \
-  if (i != num_pixels) {                                                      \
-    VP8LPredictorsSub_C[(X)](in + i, upper + i, num_pixels - i, out + i);     \
-  }                                                                           \
-}
+#define GENERATE_PREDICTOR_1(X, IN)                                         \
+  static void PredictorSub##X##_SSE2(const uint32_t* const in,              \
+                                     const uint32_t* const upper,           \
+                                     int num_pixels, uint32_t* const out) { \
+    int i;                                                                  \
+    for (i = 0; i + 4 <= num_pixels; i += 4) {                              \
+      const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);          \
+      const __m128i pred = _mm_loadu_si128((const __m128i*)&(IN));          \
+      const __m128i res = _mm_sub_epi8(src, pred);                          \
+      _mm_storeu_si128((__m128i*)&out[i], res);                             \
+    }                                                                       \
+    if (i != num_pixels) {                                                  \
+      VP8LPredictorsSub_C[(X)](in + i, WEBP_OFFSET_PTR(upper, i),           \
+                               num_pixels - i, out + i);                    \
+    }                                                                       \
+  }
 
 GENERATE_PREDICTOR_1(1, in[i - 1])       // Predictor1: L
 GENERATE_PREDICTOR_1(2, upper[i])        // Predictor2: T
diff --git a/src/enc/analysis_enc.c b/src/enc/analysis_enc.c
index 687757ae..ebb78426 100644
--- a/src/enc/analysis_enc.c
+++ b/src/enc/analysis_enc.c
@@ -126,16 +126,6 @@ static void InitHistogram(VP8Histogram* const histo) {
   histo->last_non_zero = 1;
 }
 
-static void MergeHistograms(const VP8Histogram* const in,
-                            VP8Histogram* const out) {
-  if (in->max_value > out->max_value) {
-    out->max_value = in->max_value;
-  }
-  if (in->last_non_zero > out->last_non_zero) {
-    out->last_non_zero = in->last_non_zero;
-  }
-}
-
 //------------------------------------------------------------------------------
 // Simplified k-Means, to assign Nb segments based on alpha-histogram
 
@@ -285,49 +275,6 @@ static int FastMBAnalyze(VP8EncIterator* const it) {
   return 0;
 }
 
-static int MBAnalyzeBestIntra4Mode(VP8EncIterator* const it,
-                                   int best_alpha) {
-  uint8_t modes[16];
-  const int max_mode = MAX_INTRA4_MODE;
-  int i4_alpha;
-  VP8Histogram total_histo;
-  int cur_histo = 0;
-  InitHistogram(&total_histo);
-
-  VP8IteratorStartI4(it);
-  do {
-    int mode;
-    int best_mode_alpha = DEFAULT_ALPHA;
-    VP8Histogram histos[2];
-    const uint8_t* const src = it->yuv_in_ + Y_OFF_ENC + VP8Scan[it->i4_];
-
-    VP8MakeIntra4Preds(it);
-    for (mode = 0; mode < max_mode; ++mode) {
-      int alpha;
-
-      InitHistogram(&histos[cur_histo]);
-      VP8CollectHistogram(src, it->yuv_p_ + VP8I4ModeOffsets[mode],
-                          0, 1, &histos[cur_histo]);
-      alpha = GetAlpha(&histos[cur_histo]);
-      if (IS_BETTER_ALPHA(alpha, best_mode_alpha)) {
-        best_mode_alpha = alpha;
-        modes[it->i4_] = mode;
-        cur_histo ^= 1;   // keep track of best histo so far.
-      }
-    }
-    // accumulate best histogram
-    MergeHistograms(&histos[cur_histo ^ 1], &total_histo);
-    // Note: we reuse the original samples for predictors
-  } while (VP8IteratorRotateI4(it, it->yuv_in_ + Y_OFF_ENC));
-
-  i4_alpha = GetAlpha(&total_histo);
-  if (IS_BETTER_ALPHA(i4_alpha, best_alpha)) {
-    VP8SetIntra4Mode(it, modes);
-    best_alpha = i4_alpha;
-  }
-  return best_alpha;
-}
-
 static int MBAnalyzeBestUVMode(VP8EncIterator* const it) {
   int best_alpha = DEFAULT_ALPHA;
   int smallest_alpha = 0;
@@ -371,13 +318,6 @@ static void MBAnalyze(VP8EncIterator* const it,
     best_alpha = FastMBAnalyze(it);
   } else {
     best_alpha = MBAnalyzeBestIntra16Mode(it);
-    if (enc->method_ >= 5) {
-      // We go and make a fast decision for intra4/intra16.
-      // It's usually not a good and definitive pick, but helps seeding the
-      // stats about level bit-cost.
-      // TODO(skal): improve criterion.
-      best_alpha = MBAnalyzeBestIntra4Mode(it, best_alpha);
-    }
   }
   best_uv_alpha = MBAnalyzeBestUVMode(it);
 
diff --git a/src/enc/backward_references_enc.c b/src/enc/backward_references_enc.c
index d445b40f..519b36a0 100644
--- a/src/enc/backward_references_enc.c
+++ b/src/enc/backward_references_enc.c
@@ -11,13 +11,14 @@
 //
 
 #include <assert.h>
+#include <float.h>
 #include <math.h>
 
-#include "src/enc/backward_references_enc.h"
-#include "src/enc/histogram_enc.h"
+#include "src/dsp/dsp.h"
 #include "src/dsp/lossless.h"
 #include "src/dsp/lossless_common.h"
-#include "src/dsp/dsp.h"
+#include "src/enc/backward_references_enc.h"
+#include "src/enc/histogram_enc.h"
 #include "src/utils/color_cache_utils.h"
 #include "src/utils/utils.h"
 
@@ -103,6 +104,20 @@ void VP8LBackwardRefsClear(VP8LBackwardRefs* const refs) {
   }
 }
 
+// Swaps the content of two VP8LBackwardRefs.
+static void BackwardRefsSwap(VP8LBackwardRefs* const refs1,
+                             VP8LBackwardRefs* const refs2) {
+  const int point_to_refs1 =
+      (refs1->tail_ != NULL && refs1->tail_ == &refs1->refs_);
+  const int point_to_refs2 =
+      (refs2->tail_ != NULL && refs2->tail_ == &refs2->refs_);
+  const VP8LBackwardRefs tmp = *refs1;
+  *refs1 = *refs2;
+  *refs2 = tmp;
+  if (point_to_refs2) refs1->tail_ = &refs1->refs_;
+  if (point_to_refs1) refs2->tail_ = &refs2->refs_;
+}
+
 void VP8LBackwardRefsInit(VP8LBackwardRefs* const refs, int block_size) {
   assert(refs != NULL);
   memset(refs, 0, sizeof(*refs));
@@ -154,6 +169,22 @@ static PixOrCopyBlock* BackwardRefsNewBlock(VP8LBackwardRefs* const refs) {
   return b;
 }
 
+// Return 1 on success, 0 on error.
+static int BackwardRefsClone(const VP8LBackwardRefs* const from,
+                             VP8LBackwardRefs* const to) {
+  const PixOrCopyBlock* block_from = from->refs_;
+  VP8LClearBackwardRefs(to);
+  while (block_from != NULL) {
+    PixOrCopyBlock* const block_to = BackwardRefsNewBlock(to);
+    if (block_to == NULL) return 0;
+    memcpy(block_to->start_, block_from->start_,
+           block_from->size_ * sizeof(PixOrCopy));
+    block_to->size_ = block_from->size_;
+    block_from = block_from->next_;
+  }
+  return 1;
+}
+
 extern void VP8LBackwardRefsCursorAdd(VP8LBackwardRefs* const refs,
                                       const PixOrCopy v);
 void VP8LBackwardRefsCursorAdd(VP8LBackwardRefs* const refs,
@@ -753,12 +784,18 @@ static int CalculateBestCacheSize(const uint32_t* argb, int quality,
         }
       }
     } else {
+      int code, extra_bits, extra_bits_value;
       // We should compute the contribution of the (distance,length)
       // histograms but those are the same independently from the cache size.
       // As those constant contributions are in the end added to the other
-      // histogram contributions, we can safely ignore them.
+      // histogram contributions, we can ignore them, except for the length
+      // prefix that is part of the literal_ histogram.
       int len = PixOrCopyLength(v);
       uint32_t argb_prev = *argb ^ 0xffffffffu;
+      VP8LPrefixEncode(len, &code, &extra_bits, &extra_bits_value);
+      for (i = 0; i <= cache_bits_max; ++i) {
+        ++histos[i]->literal_[NUM_LITERAL_CODES + code];
+      }
       // Update the color caches.
       do {
         if (*argb != argb_prev) {
@@ -842,16 +879,21 @@ extern int VP8LBackwardReferencesTraceBackwards(
     int xsize, int ysize, const uint32_t* const argb, int cache_bits,
     const VP8LHashChain* const hash_chain,
     const VP8LBackwardRefs* const refs_src, VP8LBackwardRefs* const refs_dst);
-static VP8LBackwardRefs* GetBackwardReferences(
-    int width, int height, const uint32_t* const argb, int quality,
-    int lz77_types_to_try, int* const cache_bits,
-    const VP8LHashChain* const hash_chain, VP8LBackwardRefs* best,
-    VP8LBackwardRefs* worst) {
-  const int cache_bits_initial = *cache_bits;
-  double bit_cost_best = -1;
+static int GetBackwardReferences(int width, int height,
+                                 const uint32_t* const argb, int quality,
+                                 int lz77_types_to_try, int cache_bits_max,
+                                 int do_no_cache,
+                                 const VP8LHashChain* const hash_chain,
+                                 VP8LBackwardRefs* const refs,
+                                 int* const cache_bits_best) {
   VP8LHistogram* histo = NULL;
-  int lz77_type, lz77_type_best = 0;
+  int i, lz77_type;
+  // Index 0 is for a color cache, index 1 for no cache (if needed).
+  int lz77_types_best[2] = {0, 0};
+  double bit_costs_best[2] = {DBL_MAX, DBL_MAX};
   VP8LHashChain hash_chain_box;
+  VP8LBackwardRefs* const refs_tmp = &refs[do_no_cache ? 2 : 1];
+  int status = 0;
   memset(&hash_chain_box, 0, sizeof(hash_chain_box));
 
   histo = VP8LAllocateHistogram(MAX_COLOR_CACHE_BITS);
@@ -860,86 +902,129 @@ static VP8LBackwardRefs* GetBackwardReferences(
   for (lz77_type = 1; lz77_types_to_try;
        lz77_types_to_try &= ~lz77_type, lz77_type <<= 1) {
     int res = 0;
-    double bit_cost;
-    int cache_bits_tmp = cache_bits_initial;
+    double bit_cost = 0.;
     if ((lz77_types_to_try & lz77_type) == 0) continue;
     switch (lz77_type) {
       case kLZ77RLE:
-        res = BackwardReferencesRle(width, height, argb, 0, worst);
+        res = BackwardReferencesRle(width, height, argb, 0, refs_tmp);
         break;
       case kLZ77Standard:
         // Compute LZ77 with no cache (0 bits), as the ideal LZ77 with a color
         // cache is not that different in practice.
-        res = BackwardReferencesLz77(width, height, argb, 0, hash_chain, worst);
+        res = BackwardReferencesLz77(width, height, argb, 0, hash_chain,
+                                     refs_tmp);
         break;
       case kLZ77Box:
         if (!VP8LHashChainInit(&hash_chain_box, width * height)) goto Error;
         res = BackwardReferencesLz77Box(width, height, argb, 0, hash_chain,
-                                        &hash_chain_box, worst);
+                                        &hash_chain_box, refs_tmp);
         break;
       default:
         assert(0);
     }
     if (!res) goto Error;
 
-    // Next, try with a color cache and update the references.
-    if (!CalculateBestCacheSize(argb, quality, worst, &cache_bits_tmp)) {
-      goto Error;
-    }
-    if (cache_bits_tmp > 0) {
-      if (!BackwardRefsWithLocalCache(argb, cache_bits_tmp, worst)) {
-        goto Error;
+    // Start with the no color cache case.
+    for (i = 1; i >= 0; --i) {
+      int cache_bits = (i == 1) ? 0 : cache_bits_max;
+
+      if (i == 1 && !do_no_cache) continue;
+
+      if (i == 0) {
+        // Try with a color cache.
+        if (!CalculateBestCacheSize(argb, quality, refs_tmp, &cache_bits)) {
+          goto Error;
+        }
+        if (cache_bits > 0) {
+          if (!BackwardRefsWithLocalCache(argb, cache_bits, refs_tmp)) {
+            goto Error;
+          }
+        }
       }
-    }
 
-    // Keep the best backward references.
-    VP8LHistogramCreate(histo, worst, cache_bits_tmp);
-    bit_cost = VP8LHistogramEstimateBits(histo);
-    if (lz77_type_best == 0 || bit_cost < bit_cost_best) {
-      VP8LBackwardRefs* const tmp = worst;
-      worst = best;
-      best = tmp;
-      bit_cost_best = bit_cost;
-      *cache_bits = cache_bits_tmp;
-      lz77_type_best = lz77_type;
+      if (i == 0 && do_no_cache && cache_bits == 0) {
+        // No need to re-compute bit_cost as it was computed at i == 1.
+      } else {
+        VP8LHistogramCreate(histo, refs_tmp, cache_bits);
+        bit_cost = VP8LHistogramEstimateBits(histo);
+      }
+
+      if (bit_cost < bit_costs_best[i]) {
+        if (i == 1) {
+          // Do not swap as the full cache analysis would have the wrong
+          // VP8LBackwardRefs to start with.
+          if (!BackwardRefsClone(refs_tmp, &refs[1])) goto Error;
+        } else {
+          BackwardRefsSwap(refs_tmp, &refs[0]);
+        }
+        bit_costs_best[i] = bit_cost;
+        lz77_types_best[i] = lz77_type;
+        if (i == 0) *cache_bits_best = cache_bits;
+      }
     }
   }
-  assert(lz77_type_best > 0);
+  assert(lz77_types_best[0] > 0);
+  assert(!do_no_cache || lz77_types_best[1] > 0);
 
   // Improve on simple LZ77 but only for high quality (TraceBackwards is
   // costly).
-  if ((lz77_type_best == kLZ77Standard || lz77_type_best == kLZ77Box) &&
-      quality >= 25) {
-    const VP8LHashChain* const hash_chain_tmp =
-        (lz77_type_best == kLZ77Standard) ? hash_chain : &hash_chain_box;
-    if (VP8LBackwardReferencesTraceBackwards(width, height, argb, *cache_bits,
-                                             hash_chain_tmp, best, worst)) {
-      double bit_cost_trace;
-      VP8LHistogramCreate(histo, worst, *cache_bits);
-      bit_cost_trace = VP8LHistogramEstimateBits(histo);
-      if (bit_cost_trace < bit_cost_best) best = worst;
+  for (i = 1; i >= 0; --i) {
+    if (i == 1 && !do_no_cache) continue;
+    if ((lz77_types_best[i] == kLZ77Standard ||
+         lz77_types_best[i] == kLZ77Box) &&
+        quality >= 25) {
+      const VP8LHashChain* const hash_chain_tmp =
+          (lz77_types_best[i] == kLZ77Standard) ? hash_chain : &hash_chain_box;
+      const int cache_bits = (i == 1) ? 0 : *cache_bits_best;
+      if (VP8LBackwardReferencesTraceBackwards(width, height, argb, cache_bits,
+                                               hash_chain_tmp, &refs[i],
+                                               refs_tmp)) {
+        double bit_cost_trace;
+        VP8LHistogramCreate(histo, refs_tmp, cache_bits);
+        bit_cost_trace = VP8LHistogramEstimateBits(histo);
+        if (bit_cost_trace < bit_costs_best[i]) {
+          BackwardRefsSwap(refs_tmp, &refs[i]);
+        }
+      }
     }
-  }
 
-  BackwardReferences2DLocality(width, best);
+    BackwardReferences2DLocality(width, &refs[i]);
+
+    if (i == 1 && lz77_types_best[0] == lz77_types_best[1] &&
+        *cache_bits_best == 0) {
+      // If the best cache size is 0 and we have the same best LZ77, just copy
+      // the data over and stop here.
+      if (!BackwardRefsClone(&refs[1], &refs[0])) goto Error;
+      break;
+    }
+  }
+  status = 1;
 
 Error:
   VP8LHashChainClear(&hash_chain_box);
   VP8LFreeHistogram(histo);
-  return best;
+  return status;
 }
 
-VP8LBackwardRefs* VP8LGetBackwardReferences(
+WebPEncodingError VP8LGetBackwardReferences(
     int width, int height, const uint32_t* const argb, int quality,
-    int low_effort, int lz77_types_to_try, int* const cache_bits,
-    const VP8LHashChain* const hash_chain, VP8LBackwardRefs* const refs_tmp1,
-    VP8LBackwardRefs* const refs_tmp2) {
+    int low_effort, int lz77_types_to_try, int cache_bits_max, int do_no_cache,
+    const VP8LHashChain* const hash_chain, VP8LBackwardRefs* const refs,
+    int* const cache_bits_best) {
   if (low_effort) {
-    return GetBackwardReferencesLowEffort(width, height, argb, cache_bits,
-                                          hash_chain, refs_tmp1);
+    VP8LBackwardRefs* refs_best;
+    *cache_bits_best = cache_bits_max;
+    refs_best = GetBackwardReferencesLowEffort(
+        width, height, argb, cache_bits_best, hash_chain, refs);
+    if (refs_best == NULL) return VP8_ENC_ERROR_OUT_OF_MEMORY;
+    // Set it in first position.
+    BackwardRefsSwap(refs_best, &refs[0]);
   } else {
-    return GetBackwardReferences(width, height, argb, quality,
-                                 lz77_types_to_try, cache_bits, hash_chain,
-                                 refs_tmp1, refs_tmp2);
+    if (!GetBackwardReferences(width, height, argb, quality, lz77_types_to_try,
+                               cache_bits_max, do_no_cache, hash_chain, refs,
+                               cache_bits_best)) {
+      return VP8_ENC_ERROR_OUT_OF_MEMORY;
+    }
   }
+  return VP8_ENC_OK;
 }
diff --git a/src/enc/backward_references_enc.h b/src/enc/backward_references_enc.h
index 103ddfdc..4c0267b4 100644
--- a/src/enc/backward_references_enc.h
+++ b/src/enc/backward_references_enc.h
@@ -16,6 +16,7 @@
 #include <assert.h>
 #include <stdlib.h>
 #include "src/webp/types.h"
+#include "src/webp/encode.h"
 #include "src/webp/format_constants.h"
 
 #ifdef __cplusplus
@@ -218,14 +219,19 @@ enum VP8LLZ77Type {
 // Evaluates best possible backward references for specified quality.
 // The input cache_bits to 'VP8LGetBackwardReferences' sets the maximum cache
 // bits to use (passing 0 implies disabling the local color cache).
-// The optimal cache bits is evaluated and set for the *cache_bits parameter.
-// The return value is the pointer to the best of the two backward refs viz,
-// refs[0] or refs[1].
-VP8LBackwardRefs* VP8LGetBackwardReferences(
+// The optimal cache bits is evaluated and set for the *cache_bits_best
+// parameter with the matching refs_best.
+// If do_no_cache == 0, refs is an array of 2 values and the best
+// VP8LBackwardRefs is put in the first element.
+// If do_no_cache != 0, refs is an array of 3 values and the best
+// VP8LBackwardRefs is put in the first element, the best value with no-cache in
+// the second element.
+// In both cases, the last element is used as temporary internally.
+WebPEncodingError VP8LGetBackwardReferences(
     int width, int height, const uint32_t* const argb, int quality,
-    int low_effort, int lz77_types_to_try, int* const cache_bits,
-    const VP8LHashChain* const hash_chain, VP8LBackwardRefs* const refs_tmp1,
-    VP8LBackwardRefs* const refs_tmp2);
+    int low_effort, int lz77_types_to_try, int cache_bits_max, int do_no_cache,
+    const VP8LHashChain* const hash_chain, VP8LBackwardRefs* const refs,
+    int* const cache_bits_best);
 
 #ifdef __cplusplus
 }
diff --git a/src/enc/config_enc.c b/src/enc/config_enc.c
index 9d482897..3518b414 100644
--- a/src/enc/config_enc.c
+++ b/src/enc/config_enc.c
@@ -39,6 +39,8 @@ int WebPConfigInitInternal(WebPConfig* config,
   config->partitions = 0;
   config->segments = 4;
   config->pass = 1;
+  config->qmin = 0;
+  config->qmax = 100;
   config->show_compressed = 0;
   config->preprocessing = 0;
   config->autofilter = 0;
@@ -106,6 +108,9 @@ int WebPValidateConfig(const WebPConfig* config) {
   if (config->filter_type < 0 || config->filter_type > 1) return 0;
   if (config->autofilter < 0 || config->autofilter > 1) return 0;
   if (config->pass < 1 || config->pass > 10) return 0;
+  if (config->qmin < 0 || config->qmax > 100 || config->qmin > config->qmax) {
+    return 0;
+  }
   if (config->show_compressed < 0 || config->show_compressed > 1) return 0;
   if (config->preprocessing < 0 || config->preprocessing > 7) return 0;
   if (config->partitions < 0 || config->partitions > 3) return 0;
diff --git a/src/enc/frame_enc.c b/src/enc/frame_enc.c
index 1aec376e..af538d83 100644
--- a/src/enc/frame_enc.c
+++ b/src/enc/frame_enc.c
@@ -31,10 +31,15 @@
 // we allow 2k of extra head-room in PARTITION0 limit.
 #define PARTITION0_SIZE_LIMIT ((VP8_MAX_PARTITION0_SIZE - 2048ULL) << 11)
 
+static float Clamp(float v, float min, float max) {
+  return (v < min) ? min : (v > max) ? max : v;
+}
+
 typedef struct {  // struct for organizing convergence in either size or PSNR
   int is_first;
   float dq;
   float q, last_q;
+  float qmin, qmax;
   double value, last_value;   // PSNR or size
   double target;
   int do_size_search;
@@ -47,7 +52,9 @@ static int InitPassStats(const VP8Encoder* const enc, PassStats* const s) {
 
   s->is_first = 1;
   s->dq = 10.f;
-  s->q = s->last_q = enc->config_->quality;
+  s->qmin = 1.f * enc->config_->qmin;
+  s->qmax = 1.f * enc->config_->qmax;
+  s->q = s->last_q = Clamp(enc->config_->quality, s->qmin, s->qmax);
   s->target = do_size_search ? (double)target_size
             : (target_PSNR > 0.) ? target_PSNR
             : 40.;   // default, just in case
@@ -56,10 +63,6 @@ static int InitPassStats(const VP8Encoder* const enc, PassStats* const s) {
   return do_size_search;
 }
 
-static float Clamp(float v, float min, float max) {
-  return (v < min) ? min : (v > max) ? max : v;
-}
-
 static float ComputeNextQ(PassStats* const s) {
   float dq;
   if (s->is_first) {
@@ -75,7 +78,7 @@ static float ComputeNextQ(PassStats* const s) {
   s->dq = Clamp(dq, -30.f, 30.f);
   s->last_q = s->q;
   s->last_value = s->value;
-  s->q = Clamp(s->q + s->dq, 0.f, 100.f);
+  s->q = Clamp(s->q + s->dq, s->qmin, s->qmax);
   return s->q;
 }
 
@@ -848,9 +851,10 @@ int VP8EncTokenLoop(VP8Encoder* const enc) {
     }
 
 #if (DEBUG_SEARCH > 0)
-    printf("#%2d metric:%.1lf -> %.1lf   last_q=%.2lf q=%.2lf dq=%.2lf\n",
+    printf("#%2d metric:%.1lf -> %.1lf   last_q=%.2lf q=%.2lf dq=%.2lf "
+           " range:[%.1f, %.1f]\n",
            num_pass_left, stats.last_value, stats.value,
-           stats.last_q, stats.q, stats.dq);
+           stats.last_q, stats.q, stats.dq, stats.qmin, stats.qmax);
 #endif
     if (enc->max_i4_header_bits_ > 0 && size_p0 > PARTITION0_SIZE_LIMIT) {
       ++num_pass_left;
diff --git a/src/enc/histogram_enc.c b/src/enc/histogram_enc.c
index a4e6bf3a..edc6e4fa 100644
--- a/src/enc/histogram_enc.c
+++ b/src/enc/histogram_enc.c
@@ -208,6 +208,7 @@ void VP8LHistogramAddSinglePixOrCopy(VP8LHistogram* const histo,
   } else if (PixOrCopyIsCacheIdx(v)) {
     const int literal_ix =
         NUM_LITERAL_CODES + NUM_LENGTH_CODES + PixOrCopyCacheIdx(v);
+    assert(histo->palette_code_bits_ != 0);
     ++histo->literal_[literal_ix];
   } else {
     int code, extra_bits;
diff --git a/src/enc/picture_tools_enc.c b/src/enc/picture_tools_enc.c
index d0e8a495..5cc25c5f 100644
--- a/src/enc/picture_tools_enc.c
+++ b/src/enc/picture_tools_enc.c
@@ -83,6 +83,19 @@ static int SmoothenBlock(const uint8_t* a_ptr, int a_stride, uint8_t* y_ptr,
   return (count == 0);
 }
 
+void WebPReplaceTransparentPixels(WebPPicture* pic, uint32_t color) {
+  if (pic != NULL && pic->use_argb) {
+    int y = pic->height;
+    uint32_t* argb = pic->argb;
+    color &= 0xffffffu;   // force alpha=0
+    WebPInitAlphaProcessing();
+    while (y-- > 0) {
+      WebPAlphaReplace(argb, pic->width, color);
+      argb += pic->argb_stride;
+    }
+  }
+}
+
 void WebPCleanupTransparentArea(WebPPicture* pic) {
   int x, y, w, h;
   if (pic == NULL) return;
@@ -165,24 +178,6 @@ void WebPCleanupTransparentArea(WebPPicture* pic) {
 #undef SIZE
 #undef SIZE2
 
-void WebPCleanupTransparentAreaLossless(WebPPicture* const pic) {
-  int x, y, w, h;
-  uint32_t* argb;
-  assert(pic != NULL && pic->use_argb);
-  w = pic->width;
-  h = pic->height;
-  argb = pic->argb;
-
-  for (y = 0; y < h; ++y) {
-    for (x = 0; x < w; ++x) {
-      if ((argb[x] & 0xff000000) == 0) {
-        argb[x] = 0x00000000;
-      }
-    }
-    argb += pic->argb_stride;
-  }
-}
-
 //------------------------------------------------------------------------------
 // Blend color and remove transparency info
 
diff --git a/src/enc/vp8i_enc.h b/src/enc/vp8i_enc.h
index fedcaeea..fc3b73be 100644
--- a/src/enc/vp8i_enc.h
+++ b/src/enc/vp8i_enc.h
@@ -505,9 +505,9 @@ int WebPPictureAllocARGB(WebPPicture* const picture, int width, int height);
 // Returns false in case of error (invalid param, out-of-memory).
 int WebPPictureAllocYUVA(WebPPicture* const picture, int width, int height);
 
-// Clean-up the RGB samples under fully transparent area, to help lossless
-// compressibility (no guarantee, though). Assumes that pic->use_argb is true.
-void WebPCleanupTransparentAreaLossless(WebPPicture* const pic);
+// Replace samples that are fully transparent by 'color' to help compressibility
+// (no guarantee, though). Assumes pic->use_argb is true.
+void WebPReplaceTransparentPixels(WebPPicture* const pic, uint32_t color);
 
 //------------------------------------------------------------------------------
 
diff --git a/src/enc/vp8l_enc.c b/src/enc/vp8l_enc.c
index 2efd403f..0b44ebe4 100644
--- a/src/enc/vp8l_enc.c
+++ b/src/enc/vp8l_enc.c
@@ -144,7 +144,8 @@ typedef enum {
   kSubGreen = 2,
   kSpatialSubGreen = 3,
   kPalette = 4,
-  kNumEntropyIx = 5
+  kPaletteAndSpatial = 5,
+  kNumEntropyIx = 6
 } EntropyIx;
 
 typedef enum {
@@ -354,11 +355,15 @@ static int GetTransformBits(int method, int histo_bits) {
 }
 
 // Set of parameters to be used in each iteration of the cruncher.
-#define CRUNCH_CONFIGS_LZ77_MAX 2
+#define CRUNCH_SUBCONFIGS_MAX 2
+typedef struct {
+  int lz77_;
+  int do_no_cache_;
+} CrunchSubConfig;
 typedef struct {
   int entropy_idx_;
-  int lz77s_types_to_try_[CRUNCH_CONFIGS_LZ77_MAX];
-  int lz77s_types_to_try_size_;
+  CrunchSubConfig sub_configs_[CRUNCH_SUBCONFIGS_MAX];
+  int sub_configs_size_;
 } CrunchConfig;
 
 #define CRUNCH_CONFIGS_MAX kNumEntropyIx
@@ -376,6 +381,9 @@ static int EncoderAnalyze(VP8LEncoder* const enc,
   int i;
   int use_palette;
   int n_lz77s;
+  // If set to 0, analyze the cache with the computed cache value. If 1, also
+  // analyze with no-cache.
+  int do_no_cache = 0;
   assert(pic != NULL && pic->argb != NULL);
 
   use_palette =
@@ -402,10 +410,13 @@ static int EncoderAnalyze(VP8LEncoder* const enc,
       return 0;
     }
     if (method == 6 && config->quality == 100) {
+      do_no_cache = 1;
       // Go brute force on all transforms.
       *crunch_configs_size = 0;
       for (i = 0; i < kNumEntropyIx; ++i) {
-        if (i != kPalette || use_palette) {
+        // We can only apply kPalette or kPaletteAndSpatial if we can indeed use
+        // a palette.
+        if ((i != kPalette && i != kPaletteAndSpatial) || use_palette) {
           assert(*crunch_configs_size < CRUNCH_CONFIGS_MAX);
           crunch_configs[(*crunch_configs_size)++].entropy_idx_ = i;
         }
@@ -414,17 +425,28 @@ static int EncoderAnalyze(VP8LEncoder* const enc,
       // Only choose the guessed best transform.
       *crunch_configs_size = 1;
       crunch_configs[0].entropy_idx_ = min_entropy_ix;
+      if (config->quality >= 75 && method == 5) {
+        // Test with and without color cache.
+        do_no_cache = 1;
+        // If we have a palette, also check in combination with spatial.
+        if (min_entropy_ix == kPalette) {
+          *crunch_configs_size = 2;
+          crunch_configs[1].entropy_idx_ = kPaletteAndSpatial;
+        }
+      }
     }
   }
   // Fill in the different LZ77s.
-  assert(n_lz77s <= CRUNCH_CONFIGS_LZ77_MAX);
+  assert(n_lz77s <= CRUNCH_SUBCONFIGS_MAX);
   for (i = 0; i < *crunch_configs_size; ++i) {
     int j;
     for (j = 0; j < n_lz77s; ++j) {
-      crunch_configs[i].lz77s_types_to_try_[j] =
+      assert(j < CRUNCH_SUBCONFIGS_MAX);
+      crunch_configs[i].sub_configs_[j].lz77_ =
           (j == 0) ? kLZ77Standard | kLZ77RLE : kLZ77Box;
+      crunch_configs[i].sub_configs_[j].do_no_cache_ = do_no_cache;
     }
-    crunch_configs[i].lz77s_types_to_try_size_ = n_lz77s;
+    crunch_configs[i].sub_configs_size_ = n_lz77s;
   }
   return 1;
 }
@@ -440,7 +462,7 @@ static int EncoderInit(VP8LEncoder* const enc) {
   int i;
   if (!VP8LHashChainInit(&enc->hash_chain_, pix_cnt)) return 0;
 
-  for (i = 0; i < 3; ++i) VP8LBackwardRefsInit(&enc->refs_[i], refs_block_size);
+  for (i = 0; i < 4; ++i) VP8LBackwardRefsInit(&enc->refs_[i], refs_block_size);
 
   return 1;
 }
@@ -769,13 +791,10 @@ static WebPEncodingError StoreImageToBitMask(
 }
 
 // Special case of EncodeImageInternal() for cache-bits=0, histo_bits=31
-static WebPEncodingError EncodeImageNoHuffman(VP8LBitWriter* const bw,
-                                              const uint32_t* const argb,
-                                              VP8LHashChain* const hash_chain,
-                                              VP8LBackwardRefs* const refs_tmp1,
-                                              VP8LBackwardRefs* const refs_tmp2,
-                                              int width, int height,
-                                              int quality, int low_effort) {
+static WebPEncodingError EncodeImageNoHuffman(
+    VP8LBitWriter* const bw, const uint32_t* const argb,
+    VP8LHashChain* const hash_chain, VP8LBackwardRefs* const refs_array,
+    int width, int height, int quality, int low_effort) {
   int i;
   int max_tokens = 0;
   WebPEncodingError err = VP8_ENC_OK;
@@ -798,13 +817,11 @@ static WebPEncodingError EncodeImageNoHuffman(VP8LBitWriter* const bw,
     err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
-  refs = VP8LGetBackwardReferences(width, height, argb, quality, 0,
-                                   kLZ77Standard | kLZ77RLE, &cache_bits,
-                                   hash_chain, refs_tmp1, refs_tmp2);
-  if (refs == NULL) {
-    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
-    goto Error;
-  }
+  err = VP8LGetBackwardReferences(
+      width, height, argb, quality, /*low_effort=*/0, kLZ77Standard | kLZ77RLE,
+      cache_bits, /*do_no_cache=*/0, hash_chain, refs_array, &cache_bits);
+  if (err != VP8_ENC_OK) goto Error;
+  refs = &refs_array[0];
   histogram_image = VP8LAllocateHistogramSet(1, cache_bits);
   if (histogram_image == NULL) {
     err = VP8_ENC_ERROR_OUT_OF_MEMORY;
@@ -860,11 +877,11 @@ static WebPEncodingError EncodeImageNoHuffman(VP8LBitWriter* const bw,
 
 static WebPEncodingError EncodeImageInternal(
     VP8LBitWriter* const bw, const uint32_t* const argb,
-    VP8LHashChain* const hash_chain, VP8LBackwardRefs refs_array[3], int width,
+    VP8LHashChain* const hash_chain, VP8LBackwardRefs refs_array[4], int width,
     int height, int quality, int low_effort, int use_cache,
     const CrunchConfig* const config, int* cache_bits, int histogram_bits,
     size_t init_byte_position, int* const hdr_size, int* const data_size) {
-  WebPEncodingError err = VP8_ENC_OK;
+  WebPEncodingError err = VP8_ENC_ERROR_OUT_OF_MEMORY;
   const uint32_t histogram_image_xysize =
       VP8LSubSampleSize(width, histogram_bits) *
       VP8LSubSampleSize(height, histogram_bits);
@@ -876,103 +893,103 @@ static WebPEncodingError EncodeImageInternal(
       3ULL * CODE_LENGTH_CODES, sizeof(*huff_tree));
   HuffmanTreeToken* tokens = NULL;
   HuffmanTreeCode* huffman_codes = NULL;
-  VP8LBackwardRefs* refs_best;
-  VP8LBackwardRefs* refs_tmp;
   uint16_t* const histogram_symbols =
       (uint16_t*)WebPSafeMalloc(histogram_image_xysize,
                                 sizeof(*histogram_symbols));
-  int lz77s_idx;
+  int sub_configs_idx;
+  int cache_bits_init, write_histogram_image;
   VP8LBitWriter bw_init = *bw, bw_best;
   int hdr_size_tmp;
+  VP8LHashChain hash_chain_histogram;  // histogram image hash chain
+  size_t bw_size_best = ~(size_t)0;
   assert(histogram_bits >= MIN_HUFFMAN_BITS);
   assert(histogram_bits <= MAX_HUFFMAN_BITS);
   assert(hdr_size != NULL);
   assert(data_size != NULL);
 
-  if (histogram_symbols == NULL) {
-    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
+  // Make sure we can allocate the different objects.
+  memset(&hash_chain_histogram, 0, sizeof(hash_chain_histogram));
+  if (huff_tree == NULL || histogram_symbols == NULL ||
+      !VP8LHashChainInit(&hash_chain_histogram, histogram_image_xysize) ||
+      !VP8LHashChainFill(hash_chain, quality, argb, width, height,
+                         low_effort)) {
     goto Error;
   }
-
   if (use_cache) {
     // If the value is different from zero, it has been set during the
     // palette analysis.
-    if (*cache_bits == 0) *cache_bits = MAX_COLOR_CACHE_BITS;
+    cache_bits_init = (*cache_bits == 0) ? MAX_COLOR_CACHE_BITS : *cache_bits;
   } else {
-    *cache_bits = 0;
+    cache_bits_init = 0;
   }
-  // 'best_refs' is the reference to the best backward refs and points to one
-  // of refs_array[0] or refs_array[1].
-  // Calculate backward references from ARGB image.
-  if (huff_tree == NULL ||
-      !VP8LHashChainFill(hash_chain, quality, argb, width, height,
-                         low_effort) ||
-      !VP8LBitWriterInit(&bw_best, 0) ||
-      (config->lz77s_types_to_try_size_ > 1 &&
+  // If several iterations will happen, clone into bw_best.
+  if (!VP8LBitWriterInit(&bw_best, 0) ||
+      ((config->sub_configs_size_ > 1 ||
+        config->sub_configs_[0].do_no_cache_) &&
        !VP8LBitWriterClone(bw, &bw_best))) {
-    err = VP8_ENC_ERROR_OUT_OF_MEMORY;
     goto Error;
   }
-  for (lz77s_idx = 0; lz77s_idx < config->lz77s_types_to_try_size_;
-       ++lz77s_idx) {
-    refs_best = VP8LGetBackwardReferences(
-        width, height, argb, quality, low_effort,
-        config->lz77s_types_to_try_[lz77s_idx], cache_bits, hash_chain,
-        &refs_array[0], &refs_array[1]);
-    if (refs_best == NULL) {
-      err = VP8_ENC_ERROR_OUT_OF_MEMORY;
-      goto Error;
-    }
-    // Keep the best references aside and use the other element from the first
-    // two as a temporary for later usage.
-    refs_tmp = &refs_array[refs_best == &refs_array[0] ? 1 : 0];
-
-    histogram_image =
-        VP8LAllocateHistogramSet(histogram_image_xysize, *cache_bits);
-    tmp_histo = VP8LAllocateHistogram(*cache_bits);
-    if (histogram_image == NULL || tmp_histo == NULL) {
-      err = VP8_ENC_ERROR_OUT_OF_MEMORY;
-      goto Error;
-    }
-
-    // Build histogram image and symbols from backward references.
-    if (!VP8LGetHistoImageSymbols(width, height, refs_best, quality, low_effort,
-                                  histogram_bits, *cache_bits, histogram_image,
-                                  tmp_histo, histogram_symbols)) {
-      err = VP8_ENC_ERROR_OUT_OF_MEMORY;
-      goto Error;
-    }
-    // Create Huffman bit lengths and codes for each histogram image.
-    histogram_image_size = histogram_image->size;
-    bit_array_size = 5 * histogram_image_size;
-    huffman_codes = (HuffmanTreeCode*)WebPSafeCalloc(bit_array_size,
-                                                     sizeof(*huffman_codes));
-    // Note: some histogram_image entries may point to tmp_histos[], so the
-    // latter need to outlive the following call to GetHuffBitLengthsAndCodes().
-    if (huffman_codes == NULL ||
-        !GetHuffBitLengthsAndCodes(histogram_image, huffman_codes)) {
-      err = VP8_ENC_ERROR_OUT_OF_MEMORY;
-      goto Error;
-    }
-    // Free combined histograms.
-    VP8LFreeHistogramSet(histogram_image);
-    histogram_image = NULL;
-
-    // Free scratch histograms.
-    VP8LFreeHistogram(tmp_histo);
-    tmp_histo = NULL;
+  for (sub_configs_idx = 0; sub_configs_idx < config->sub_configs_size_;
+       ++sub_configs_idx) {
+    const CrunchSubConfig* const sub_config =
+        &config->sub_configs_[sub_configs_idx];
+    int cache_bits_best, i_cache;
+    err = VP8LGetBackwardReferences(width, height, argb, quality, low_effort,
+                                    sub_config->lz77_, cache_bits_init,
+                                    sub_config->do_no_cache_, hash_chain,
+                                    &refs_array[0], &cache_bits_best);
+    if (err != VP8_ENC_OK) goto Error;
 
-    // Color Cache parameters.
-    if (*cache_bits > 0) {
-      VP8LPutBits(bw, 1, 1);
-      VP8LPutBits(bw, *cache_bits, 4);
-    } else {
-      VP8LPutBits(bw, 0, 1);
-    }
+    for (i_cache = 0; i_cache < (sub_config->do_no_cache_ ? 2 : 1); ++i_cache) {
+      const int cache_bits_tmp = (i_cache == 0) ? cache_bits_best : 0;
+      // Speed-up: no need to study the no-cache case if it was already studied
+      // in i_cache == 0.
+      if (i_cache == 1 && cache_bits_best == 0) break;
+
+      // Reset the bit writer for this iteration.
+      VP8LBitWriterReset(&bw_init, bw);
+
+      // Build histogram image and symbols from backward references.
+      histogram_image =
+          VP8LAllocateHistogramSet(histogram_image_xysize, cache_bits_tmp);
+      tmp_histo = VP8LAllocateHistogram(cache_bits_tmp);
+      if (histogram_image == NULL || tmp_histo == NULL ||
+          !VP8LGetHistoImageSymbols(width, height, &refs_array[i_cache],
+                                    quality, low_effort, histogram_bits,
+                                    cache_bits_tmp, histogram_image, tmp_histo,
+                                    histogram_symbols)) {
+        goto Error;
+      }
+      // Create Huffman bit lengths and codes for each histogram image.
+      histogram_image_size = histogram_image->size;
+      bit_array_size = 5 * histogram_image_size;
+      huffman_codes = (HuffmanTreeCode*)WebPSafeCalloc(bit_array_size,
+                                                       sizeof(*huffman_codes));
+      // Note: some histogram_image entries may point to tmp_histos[], so the
+      // latter need to outlive the following call to
+      // GetHuffBitLengthsAndCodes().
+      if (huffman_codes == NULL ||
+          !GetHuffBitLengthsAndCodes(histogram_image, huffman_codes)) {
+        goto Error;
+      }
+      // Free combined histograms.
+      VP8LFreeHistogramSet(histogram_image);
+      histogram_image = NULL;
+
+      // Free scratch histograms.
+      VP8LFreeHistogram(tmp_histo);
+      tmp_histo = NULL;
+
+      // Color Cache parameters.
+      if (cache_bits_tmp > 0) {
+        VP8LPutBits(bw, 1, 1);
+        VP8LPutBits(bw, cache_bits_tmp, 4);
+      } else {
+        VP8LPutBits(bw, 0, 1);
+      }
 
-    // Huffman image + meta huffman.
-    {
-      const int write_histogram_image = (histogram_image_size > 1);
+      // Huffman image + meta huffman.
+      write_histogram_image = (histogram_image_size > 1);
       VP8LPutBits(bw, write_histogram_image, 1);
       if (write_histogram_image) {
         uint32_t* const histogram_argb =
@@ -980,10 +997,7 @@ static WebPEncodingError EncodeImageInternal(
                                       sizeof(*histogram_argb));
         int max_index = 0;
         uint32_t i;
-        if (histogram_argb == NULL) {
-          err = VP8_ENC_ERROR_OUT_OF_MEMORY;
-          goto Error;
-        }
+        if (histogram_argb == NULL) goto Error;
         for (i = 0; i < histogram_image_xysize; ++i) {
           const int symbol_index = histogram_symbols[i] & 0xffff;
           histogram_argb[i] = (symbol_index << 8);
@@ -995,65 +1009,64 @@ static WebPEncodingError EncodeImageInternal(
 
         VP8LPutBits(bw, histogram_bits - 2, 3);
         err = EncodeImageNoHuffman(
-            bw, histogram_argb, hash_chain, refs_tmp, &refs_array[2],
+            bw, histogram_argb, &hash_chain_histogram, &refs_array[2],
             VP8LSubSampleSize(width, histogram_bits),
             VP8LSubSampleSize(height, histogram_bits), quality, low_effort);
         WebPSafeFree(histogram_argb);
         if (err != VP8_ENC_OK) goto Error;
       }
-    }
 
-    // Store Huffman codes.
-    {
-      int i;
-      int max_tokens = 0;
-      // Find maximum number of symbols for the huffman tree-set.
-      for (i = 0; i < 5 * histogram_image_size; ++i) {
-        HuffmanTreeCode* const codes = &huffman_codes[i];
-        if (max_tokens < codes->num_symbols) {
-          max_tokens = codes->num_symbols;
+      // Store Huffman codes.
+      {
+        int i;
+        int max_tokens = 0;
+        // Find maximum number of symbols for the huffman tree-set.
+        for (i = 0; i < 5 * histogram_image_size; ++i) {
+          HuffmanTreeCode* const codes = &huffman_codes[i];
+          if (max_tokens < codes->num_symbols) {
+            max_tokens = codes->num_symbols;
+          }
+        }
+        tokens = (HuffmanTreeToken*)WebPSafeMalloc(max_tokens, sizeof(*tokens));
+        if (tokens == NULL) goto Error;
+        for (i = 0; i < 5 * histogram_image_size; ++i) {
+          HuffmanTreeCode* const codes = &huffman_codes[i];
+          StoreHuffmanCode(bw, huff_tree, tokens, codes);
+          ClearHuffmanTreeIfOnlyOneSymbol(codes);
         }
       }
-      tokens = (HuffmanTreeToken*)WebPSafeMalloc(max_tokens, sizeof(*tokens));
-      if (tokens == NULL) {
-        err = VP8_ENC_ERROR_OUT_OF_MEMORY;
-        goto Error;
+      // Store actual literals.
+      hdr_size_tmp = (int)(VP8LBitWriterNumBytes(bw) - init_byte_position);
+      err = StoreImageToBitMask(bw, width, histogram_bits, &refs_array[i_cache],
+                                histogram_symbols, huffman_codes);
+      if (err != VP8_ENC_OK) goto Error;
+      // Keep track of the smallest image so far.
+      if (VP8LBitWriterNumBytes(bw) < bw_size_best) {
+        bw_size_best = VP8LBitWriterNumBytes(bw);
+        *cache_bits = cache_bits_tmp;
+        *hdr_size = hdr_size_tmp;
+        *data_size =
+            (int)(VP8LBitWriterNumBytes(bw) - init_byte_position - *hdr_size);
+        VP8LBitWriterSwap(bw, &bw_best);
       }
-      for (i = 0; i < 5 * histogram_image_size; ++i) {
-        HuffmanTreeCode* const codes = &huffman_codes[i];
-        StoreHuffmanCode(bw, huff_tree, tokens, codes);
-        ClearHuffmanTreeIfOnlyOneSymbol(codes);
+      WebPSafeFree(tokens);
+      tokens = NULL;
+      if (huffman_codes != NULL) {
+        WebPSafeFree(huffman_codes->codes);
+        WebPSafeFree(huffman_codes);
+        huffman_codes = NULL;
       }
     }
-    // Store actual literals.
-    hdr_size_tmp = (int)(VP8LBitWriterNumBytes(bw) - init_byte_position);
-    err = StoreImageToBitMask(bw, width, histogram_bits, refs_best,
-                              histogram_symbols, huffman_codes);
-    // Keep track of the smallest image so far.
-    if (lz77s_idx == 0 ||
-        VP8LBitWriterNumBytes(bw) < VP8LBitWriterNumBytes(&bw_best)) {
-      *hdr_size = hdr_size_tmp;
-      *data_size =
-          (int)(VP8LBitWriterNumBytes(bw) - init_byte_position - *hdr_size);
-      VP8LBitWriterSwap(bw, &bw_best);
-    }
-    // Reset the bit writer for the following iteration if any.
-    if (config->lz77s_types_to_try_size_ > 1) VP8LBitWriterReset(&bw_init, bw);
-    WebPSafeFree(tokens);
-    tokens = NULL;
-    if (huffman_codes != NULL) {
-      WebPSafeFree(huffman_codes->codes);
-      WebPSafeFree(huffman_codes);
-      huffman_codes = NULL;
-    }
   }
   VP8LBitWriterSwap(bw, &bw_best);
+  err = VP8_ENC_OK;
 
  Error:
   WebPSafeFree(tokens);
   WebPSafeFree(huff_tree);
   VP8LFreeHistogramSet(histogram_image);
   VP8LFreeHistogram(tmp_histo);
+  VP8LHashChainClear(&hash_chain_histogram);
   if (huffman_codes != NULL) {
     WebPSafeFree(huffman_codes->codes);
     WebPSafeFree(huffman_codes);
@@ -1095,8 +1108,7 @@ static WebPEncodingError ApplyPredictFilter(const VP8LEncoder* const enc,
   VP8LPutBits(bw, pred_bits - 2, 3);
   return EncodeImageNoHuffman(
       bw, enc->transform_data_, (VP8LHashChain*)&enc->hash_chain_,
-      (VP8LBackwardRefs*)&enc->refs_[0],  // cast const away
-      (VP8LBackwardRefs*)&enc->refs_[1], transform_width, transform_height,
+      (VP8LBackwardRefs*)&enc->refs_[0], transform_width, transform_height,
       quality, low_effort);
 }
 
@@ -1116,8 +1128,7 @@ static WebPEncodingError ApplyCrossColorFilter(const VP8LEncoder* const enc,
   VP8LPutBits(bw, ccolor_transform_bits - 2, 3);
   return EncodeImageNoHuffman(
       bw, enc->transform_data_, (VP8LHashChain*)&enc->hash_chain_,
-      (VP8LBackwardRefs*)&enc->refs_[0],  // cast const away
-      (VP8LBackwardRefs*)&enc->refs_[1], transform_width, transform_height,
+      (VP8LBackwardRefs*)&enc->refs_[0], transform_width, transform_height,
       quality, low_effort);
 }
 
@@ -1464,8 +1475,8 @@ static WebPEncodingError EncodePalette(VP8LBitWriter* const bw, int low_effort,
   }
   tmp_palette[0] = palette[0];
   return EncodeImageNoHuffman(bw, tmp_palette, &enc->hash_chain_,
-                              &enc->refs_[0], &enc->refs_[1], palette_size, 1,
-                              20 /* quality */, low_effort);
+                              &enc->refs_[0], palette_size, 1, /*quality=*/20,
+                              low_effort);
 }
 
 // -----------------------------------------------------------------------------
@@ -1491,7 +1502,7 @@ static void VP8LEncoderDelete(VP8LEncoder* enc) {
   if (enc != NULL) {
     int i;
     VP8LHashChainClear(&enc->hash_chain_);
-    for (i = 0; i < 3; ++i) VP8LBackwardRefsClear(&enc->refs_[i]);
+    for (i = 0; i < 4; ++i) VP8LBackwardRefsClear(&enc->refs_[i]);
     ClearTransformBuffer(enc);
     WebPSafeFree(enc);
   }
@@ -1541,7 +1552,7 @@ static int EncodeStreamHook(void* input, void* data2) {
   int data_size = 0;
   int use_delta_palette = 0;
   int idx;
-  size_t best_size = 0;
+  size_t best_size = ~(size_t)0;
   VP8LBitWriter bw_init = *bw, bw_best;
   (void)data2;
 
@@ -1553,11 +1564,13 @@ static int EncodeStreamHook(void* input, void* data2) {
 
   for (idx = 0; idx < num_crunch_configs; ++idx) {
     const int entropy_idx = crunch_configs[idx].entropy_idx_;
-    enc->use_palette_ = (entropy_idx == kPalette);
+    enc->use_palette_ =
+        (entropy_idx == kPalette) || (entropy_idx == kPaletteAndSpatial);
     enc->use_subtract_green_ =
         (entropy_idx == kSubGreen) || (entropy_idx == kSpatialSubGreen);
-    enc->use_predict_ =
-        (entropy_idx == kSpatial) || (entropy_idx == kSpatialSubGreen);
+    enc->use_predict_ = (entropy_idx == kSpatial) ||
+                        (entropy_idx == kSpatialSubGreen) ||
+                        (entropy_idx == kPaletteAndSpatial);
     if (low_effort) {
       enc->use_cross_color_ = 0;
     } else {
@@ -1640,7 +1653,7 @@ static int EncodeStreamHook(void* input, void* data2) {
     if (err != VP8_ENC_OK) goto Error;
 
     // If we are better than what we already have.
-    if (idx == 0 || VP8LBitWriterNumBytes(bw) < best_size) {
+    if (VP8LBitWriterNumBytes(bw) < best_size) {
       best_size = VP8LBitWriterNumBytes(bw);
       // Store the BitWriter.
       VP8LBitWriterSwap(bw, &bw_best);
@@ -1816,7 +1829,7 @@ Error:
 }
 
 #undef CRUNCH_CONFIGS_MAX
-#undef CRUNCH_CONFIGS_LZ77_MAX
+#undef CRUNCH_SUBCONFIGS_MAX
 
 int VP8LEncodeImage(const WebPConfig* const config,
                     const WebPPicture* const picture) {
diff --git a/src/enc/vp8li_enc.h b/src/enc/vp8li_enc.h
index d2d0fc50..94210ce9 100644
--- a/src/enc/vp8li_enc.h
+++ b/src/enc/vp8li_enc.h
@@ -71,7 +71,7 @@ typedef struct {
   uint32_t palette_[MAX_PALETTE_SIZE];
 
   // Some 'scratch' (potentially large) objects.
-  struct VP8LBackwardRefs refs_[3];  // Backward Refs array for temporaries.
+  struct VP8LBackwardRefs refs_[4];  // Backward Refs array for temporaries.
   VP8LHashChain hash_chain_;         // HashChain data for constructing
                                      // backward references.
 } VP8LEncoder;
diff --git a/src/enc/webp_enc.c b/src/enc/webp_enc.c
index 9f4b10c2..ce2db2e9 100644
--- a/src/enc/webp_enc.c
+++ b/src/enc/webp_enc.c
@@ -400,7 +400,7 @@ int WebPEncode(const WebPConfig* config, WebPPicture* pic) {
     }
 
     if (!config->exact) {
-      WebPCleanupTransparentAreaLossless(pic);
+      WebPReplaceTransparentPixels(pic, 0x000000);
     }
 
     ok = VP8LEncodeImage(config, pic);  // Sets pic->error in case of problem.
diff --git a/src/utils/utils.c b/src/utils/utils.c
index 764f752b..6080e19e 100644
--- a/src/utils/utils.c
+++ b/src/utils/utils.c
@@ -231,7 +231,7 @@ void WebPFree(void* ptr) {
 void WebPCopyPlane(const uint8_t* src, int src_stride,
                    uint8_t* dst, int dst_stride, int width, int height) {
   assert(src != NULL && dst != NULL);
-  assert(src_stride >= width && dst_stride >= width);
+  assert(abs(src_stride) >= width && abs(dst_stride) >= width);
   while (height-- > 0) {
     memcpy(dst, src, width);
     src += src_stride;
diff --git a/src/webp/encode.h b/src/webp/encode.h
index 655166e7..54938dfe 100644
--- a/src/webp/encode.h
+++ b/src/webp/encode.h
@@ -20,7 +20,7 @@
 extern "C" {
 #endif
 
-#define WEBP_ENCODER_ABI_VERSION 0x020f    // MAJOR(8b) + MINOR(8b)
+#define WEBP_ENCODER_ABI_VERSION 0x0210    // MAJOR(8b) + MINOR(8b)
 
 // Note: forward declaring enumerations is not allowed in (strict) C and C++,
 // the types are left here for reference.
@@ -148,7 +148,8 @@ struct WebPConfig {
   int use_delta_palette;  // reserved for future lossless feature
   int use_sharp_yuv;      // if needed, use sharp (and slow) RGB->YUV conversion
 
-  uint32_t pad[2];        // padding for later use
+  int qmin;               // minimum permissible quality factor
+  int qmax;               // maximum permissible quality factor
 };
 
 // Enumerate some predefined settings for WebPConfig, depending on the type